Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Any recommendations for 24/96 => 16/44 => 24/96 for transparency (Read 22674 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #25 –
Quote from: MLXXX on
In another thread in this forum,  [a href='index.php?showtopic=106442']two hi rez vs redbook train wrecks in progress[/a], reference is made to a timing discrepancy between a native 24/96 file, and a version of it converted to 16/44, and then back to 24/96, provided over at AVS Forum:
Quote from: mzil on
There's no reason to cheat to pass the ABX test on their test song Mosaic, at least; I know didn't. Simply pause the music at a particularly telling location and notice the alteration due to the time misalignment. I can clearly hear the guitar slap at the end of this phrase repeated loop of this riff I posted ends with a "cha" vs "chip" sound.



Based on my experience with such things, files need to be time aligned within 1 mSec or better or there is a good chance of an audible slap or brief echo during at least some change-overs.

1 mSec is 44 samples at 44 KHz or 96 samples at 96 KHz.

The Mosaic delay problem was stated to be on the order of 10 mSec, so its a problem.

There are many ways to ruin the bit-perfect nature of files and not introduce audible artifacts.  Failing a nulling test is therefore a non-event.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #26 –
Yup, it was the DAC I am not a bat.
FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #27 –
Quote from: Arnold B. Krueger on
Failing a nulling test is therefore a non-event.

In some cases perhaps (possibly with Sonic Process, which was used at AVS), but with SoX (and recently, also with Izotope) you have a linear-phase, 0 delay, 0 gain system: if the passband doesn't null after conversion, then an error has occurred.

Quote from: Satellite_6 on
Yup, it was the DAC I am not a bat.

Good news.

MLXXX, this is salient for your tests.  You need to make sure that your participants first test their play-back chain, and you should probably provide upsampled to 192kHz versions of the test files for soundcards that only really support 192kHz well (which seems to be common these days).

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #28 –
I may have been feeling a bit cranky but
Quote
Where were you when all of these facts came to light? ;-)

this is very old news and you 've no doubt heard the aphorism about what yesterday's news and old newspapers are good for.

I understand being interested in something new (new to one's self) and it is certainly frustrating to have no one to hash it out with. It just seems amazing to me that so many people are willing to contribute, over and over again, at such great length too, to a topic that, for the great majority of people, will provide as much resolution as arguing about differing religious viewpoints.

Let me point out something that may or may not be very well known. Cool Edit has a reputation for doing very good resampling. I believe that it does. Just for the sake of discussion, I did a little test.
(1) generate a complex tone in 32/44.1-- I used the "bell" preset with the base frequency increased to 880Hz. This produced frequencies to above 6kHz. All signal drops off very sharply before 7kHz, so it is very unlikely there was any hidden ultrasonic frequencies to do sneaky tricks. Maximum peaks were a bit below -3dBfs.
(2) convert to 32/96.
(3) convert back to 32/44.1
(4) compare data – invert one 32/44.1 and add to the other 32/44.1

This reveals a very definite data difference which, while at a much lower level than the original, is quite audible without amplification. Amplified to match the original at peak level, it sounds pretty close to the original.

Zooming in on the first 28 samples of the original and the final versions (not the difference) I see that everything looks to be in the same place (i.e. the file comparison does not appear to be offset) but the sample values are very slightly different e.g. 20866.852 vs 20963.926. According to Cool Edit Statistics, the peak levels in the two files differs by 0.02dB but the RMS average is the same (to the two decimal places displayed).

I imagined I could hear a difference between the original and final versions. I put it to a test with a 15 trial ABX test. I tried to be diligent but I must admit to not being very enthusiastic. Occasionally I went back to A and B and verified that my imagined difference was still audible, but I probably did so for less than half of the trials.

Listening to the X and Y test presentations, I also imagined I could hear that slight, difference between X and Y, but I kept being unsure which matched to A, which to B. I made a best guess each time. The final results were 4 correct out of 15, which the program (WinABX) calculated at 97.7% probability of guessing (I'm not sure I remember that .7 digit correctly as I did not save the results).

The point of all this is that maybe I really did hear a difference but just did not match the X properly against A/B. I kept thinking I heard a difference but most times I was uncertain whether X matched with B and Y matched with A or vice versa, and it just seemed like too much energy expenditure to keep comparing them again and again in an attempt to be sure. Therefore I forced myself to make a choice after listening to each pair several times. Perhaps many people proceed similarly. Maybe many of them feel more certain than I did, of which matches to which, while still really having the same difficulty I had.

This could be taken as strong evidence that I could not hear a difference or it could be evidence that I just could not keep the differences straight in my short term memory. Is this a normal part of the controversy?

Personally I don't care. If the difference is too little to be certain about, I don't expect to ever notice it in listening to music. It doesn't make any difference whether it is real or not and is no justification for worrying about "hi rez". Audible distortion and noise, even when very slight, are much more important (but of course, up to a point, can be ignored when enjoying music).

While many people do seem to care about the possibility of a difference, I believe it all comes down the same way when listening to music: it doesn't make the slightest bit of difference if they are actually listening to the music vs trying to test themselves, which generally, by the nature of human consciousness, means actively deluding themselves.

By the way, performing the same data test (steps 1through 4 above) in Audition 3, the peak difference is below -120dB (vs -25.5dB with Cool Edit 2000), so the resampling calcs are not the same. Any normal resampling with Cool Edit (i.e. 96kHz to 44.1kHz or 32/44.1 to 16/44.1) is extremely likely to be transparent -- except if one reduces bit depth without dithering and the data doesn't have enough random noise to make a difference, as in a computer generated tone at 24 or 32 bit).

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #29 –
I passed one from that link in the second post. I failed the first two and then I did this and I stopped. I will try the other samples at a later date and I will also try to repeat this result.

ODAC/O2 -> Audeze LCD-2

foo_abx 1.3.4 report
foobar2000 v1.3.4
2014/11/11 19:21:20

File A: D:\FILES\2 MEDIA\AUDIO\test\3a.flac
File B: D:\FILES\2 MEDIA\AUDIO\test\3b.flac

19:21:20 : Test started.
19:21:33 : 01/01  50.0%
19:21:42 : 02/02  25.0%
19:22:04 : 03/03  12.5%
19:22:43 : 04/04  6.3%
19:23:00 : 05/05  3.1%
19:23:24 : 06/06  1.6%
19:23:54 : 07/07  0.8%
19:24:14 : 08/08  0.4%
19:24:34 : 09/09  0.2%
19:25:30 : 10/10  0.1%
19:25:36 : Test finished.

----------
Total: 10/10 (0.1%)
FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #30 –
Quote from: AndyH-ha on
This reveals a very definite data difference which, while at a much lower level than the original, is quite audible without amplification. Amplified to match the original at peak level, it sounds pretty close to the original.


IME this is a typical result of sample time shift. Subtracting the same track shifted by even a single sample (such shift can happen during two resamplings in a row) produces a weak but audible sound of the original. I am afraid such test does not reveal much meaningful.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #31 –
Quote from: AndyH-ha on
Let me point out something that may or may not be very well known. Cool Edit has a reputation for doing very good resampling. I believe that it does. Just for the sake of discussion, I did a little test.



'Cool Edit' has been obsolete for years.

They're now up to Audition v6

IIRC v3 was released 'free' in the last year or so

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #32 –
Quote from: Satellite_6 on
I passed one from that link in the second post
...
foo_abx 1.3.4 report
...
----------
Total: 10/10 (0.1%)

From the timestamps in the report, I'm guessing that you listened to only a portion of the sample.

If that's the case then it's essential that the ends of the portion are properly faded; otherwise, the test is invalid.

Does anyone know how foo_abx 1.3.4 works in this respect?

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #33 –
Quote from: krabapple on
They're now up to Audition v6  IIRC v3 was released 'free' in the last year or so
Downloading it if free, however using it without having bought the license is a different matter: " The serial numbers provided as a part of the download may only be used by customers who legitimately purchased CS2 or Acrobat 7 and need to maintain their current use of these products." -adobe

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #34 –
Quote
Subtracting the same track shifted by even a single sample (such shift can happen during two resamplings in a row) produces a weak but audible sound of the original.


I am well aware of the differing results when the data is not aligned sample to sample exactly but I see no evidence that is the case here. First, it is possible to mark a particular sample so one can be certain; I pulled the fourth sample up by 6dB before starting the converting process, which makes it easy to identify when zoomed in close. In addition, I can see that the time of each sample is the same and that the total number of samples is the same. What would you give as evidence that there has been a shift?

When  resampling, either to 96kHz or to 88.2kHz, every sample in the output is a newly calculated value. The program does not, for example, just keep the original samples and calculate new vales for additional samples to stick between them when upsampling. Likewise, when downsampling, each output sample is a new calculated value, it does not just drop samples out.

This means that the final version 44.1 samples have different values than the original 44.1 samples, as I stated, with example, in my original post.  Whether or not some other programs somehow keep the original sample values is something I don't know but there are clearly different resampling methods, as evidenced by the different results.

From what I see in my tests, the variance of a given sample value, between the original 44.1 version and the resampled back to 44.1 version, is less in Audition 3 than in CE2K, which I would guess explains the difference in the mix-paste inverted compare outputs.

Quote
'Cool Edit' has been obsolete for years.
They're now up to Audition v6
IIRC v3 was released 'free' in the last year or so


Yes, unfortunately it can no longer be purchased but it is still widely used. Its functionality has not degraded. CE2K was mentioned in regard to its excellent resampling earlier in this thread.

I can't speak to the latest Audition but from what I read, its functionality has been somewhat reduced as far as my, and many other people's, use goes: cleaning up audio recorded from LPs and cassettes. I think there has been some emphasis shifted to working with the audio of multi-media projects. Audition's user interface was definitely degraded in Audition 3, the only version I purchased, or have even seen. This results in an somewhat expensive program almost never being used on my system.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #35 –
Created 44.1/32 sweep 50Hz - 20Khz in sox. Resampled to 192kHz and back to 44.1kHz. All with -v quality, keeping int32 (sox's native format for inter-effect communication)

This is spectrum of the difference wav. To me it is inaudible. Of course quality of the resampling makes difference, not doubt about that.



Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #36 –
Quote from: bandpass on
Quote from: Satellite_6 on
I passed one from that link in the second post
...
foo_abx 1.3.4 report
...
----------
Total: 10/10 (0.1%)

From the timestamps in the report, I'm guessing that you listened to only a portion of the sample.

If that's the case then it's essential that the ends of the portion are properly faded; otherwise, the test is invalid.

Does anyone know how foo_abx 1.3.4 works in this respect?


If the ABX plugin is flawed then it is broken and no one can ever pass a valid test.

If this is the case then it needs to be removed from the foobar2000 site.

I failed with all of the other samples so the difference is certainly subtle-non-existent but it might be there.


FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #37 –
I can't be sure that foo_abx is flawed in this way, since I've never used it.  I'm surprised no-one round here knows though.

To be safe, shorter sections of audio could be prepared 'off-line' (or just listen to the whole sample).

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #38 –
Short-term auditory memory is how long?

Disallowing a testee to use the controls available and foce him to listen to an entire test subject is not reasonable (IMO).

However, prepping samples offline is a good idea if there is actually an issue.  I really should say was an issue as there is a 2.0 version of the plugin.

I haven't encountered any such tells with the version in question, though I haven't gone looking. The only forms of cheating I have ever felt inclined to do is gain-riding, boosting treble and looping extremely short portions.  This last technique would probably have worked far more often if the plugin had a gross transitioning problem.  It hasn't.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #39 –
If it doesn't fade the playback at the ends, then it has a gross transitioning problem (for its intended purpose at least).  For example, consider ABXing a 20kHz tone with silence: without fading, the ends of the tone will give an audible click—it's phase dependent to a certain extent, but even cutting at zero-crossings is un-necessarily playing with fire.  A half-cosine envelope of a few ms applied at each end is trivial to implement and eliminates the problem.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #40 –
Why don't you try it yourself and see if you can hear a click because I cannot.
FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #41 –
Quote from: AndyH-ha on
Whether or not some other programs somehow keep the original sample values is something I don't know
It's certainly possible to upsample losslessly and "perfectly" with integer upsampling. Like you, I don't know if anything does it this way.

Downsampling back to the original is a different matter. It's theoretically possible; trivial if you assume the 88.2kHz version is already band limited to 22kHz - just throw away every other sample. However, if you don't make this assumption you must include a sinc filter, and then it's confounded by numerical accuracy problems. Even then, I have a vague recollection that it's possible to "cheat" to make it work out overall, but I can't recall how.

It is however possible to design something that's "perfect" and nulls well enough in the audio band, even though you don't get exactly the same samples back.

(By perfect, I mean sinc-like to below the noise floor of the digital audio signal.)

Cheers,
David.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #42 –
Quote from: krabapple on
Quote from: AndyH-ha on
Let me point out something that may or may not be very well known. Cool Edit has a reputation for doing very good resampling. I believe that it does. Just for the sake of discussion, I did a little test.


'Cool Edit' has been obsolete for years.


Based on what?

Quote
They're now up to Audition v6


Is there any evidence that its resampling (or any other core function) is now meaningfully (audibly) more accurate?

AFAIK basic resampling has been a fairly stable technology since say, Y2K probably many years earlier.

Here's a couple of counter examples: 

(1) I just downloaded the latest greatest version of Audacity and while its accurate enough for what I wanted to do which was do full-duplex recording, many of its other functions (particularly signal analysis) were obviously flawed or deficient in many in many ways that CEP 2.1 from 2002 is not. 

(2) I tried to keep up with Audition's new releases but the ones I tested had deficiencies and omissions that CEP 2.1 did not.


Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #43 –
Quote from: bandpass on
If it doesn't fade the playback at the ends, then it has a gross transitioning problem (for its intended purpose at least).  For example, consider ABXing a 20kHz tone with silence: without fading, the ends of the tone will give an audible click—it's phase dependent to a certain extent, but even cutting at zero-crossings is un-necessarily playing with fire.  A half-cosine envelope of a few ms applied at each end is trivial to implement and eliminates the problem.


The intended purpose of a test file is that people focus their attention on the body of the file, not the split seconds of music at its ends.

One problem with fade ins and fade outs is that avid ABXers often make them the focus of their listening tests and even completely ignore the body of the files you want them to compare.

As a practical matter all or almost all  of the transitions that an honest listener will audition during an ABX test are under the control of the comparator, not the file creator.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #44 –
Quote from: greynol on
Short-term auditory memory is how long?


A few seconds, max.  My personal number is that anything under a second should be OK.

Quote
Disallowing a testee to use the controls available and force him to listen to an entire test subject is not reasonable (IMO).


Totally agreed. Every bit of evidence I know about says that the purpose of long term listening is finding critical passages that you actually focus on. Critical passages just naturally seem to be a few seconds long at the max.

Quote
However, prepping samples offline is a good idea if there is actually an issue.


I took a lot of abuse for doing that for my www.pcabx.com web site. However the abuse was AFAIK all based on philosophical grounds, not actual problems that were observed during earnest use of the samples.

When people can't hear what they've been educated to believe is easy to hear, they tend to blame the tools and the files.  They don't know that they have been mislead by false claims.

Quote
I really should say was an issue as there is a 2.0 version of the plugin.


The 2.0 version of the plugin is available where?

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #45 –
If you are claiming that I exclusively listened to the very beginning or the very end of the track then I did not, so unless you have a real criticism I think that this test is valid.

FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #46 –
Okay, I've just downloaded foobar + foo_abx 1.3.4 and tested it: it's broken.
The problem is not with the ends of the files as they are loaded in; it's with the artificial ends that are created by using the "Set start" and "Set end" buttons (which, from what folk are saying above, seem to be almost always used).  The artificial ends must be faded; with foo_abx they are not, and as a result I can easily ABX a 1 sec portion of a 3 sec file that I can't when playing the whole thing.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #48 –
Code: [Select]
foo_abx 1.3.4 report
foobar2000 v1.3.5
2014/11/13 20:58:02

File A: G:\fb\a.wav
File B: G:\fb\b.wav

20:58:02 : Test started.
20:59:40 : 01/01  50.0%
20:59:44 : 02/02  25.0%
20:59:48 : 03/03  12.5%
20:59:52 : 04/04  6.3%
20:59:56 : 05/05  3.1%
21:00:00 : 06/06  1.6%
21:00:04 : 07/07  0.8%
21:00:08 : 08/08  0.4%
21:00:12 : 09/09  0.2%
21:00:16 : 10/10  0.1%
21:00:20 : 11/11  0.0%
21:00:23 : 12/12  0.0%
21:01:09 : Test finished.

 ---------- 
Total: 12/12 (0.0%)

The files are the ones from the thought experiment I described above: 20kHz tone vs. silence.

Code: [Select]
 sox -r 44100 -n -b16 b.wav fade h .1 3 .1 gain -6 stats
 sox -r 44100 -n -b16 a.wav synth sine 20k fade h .1 3 .1 gain -6 stats


The new version of the ABX component behaves differently: at the the leading edge ("Set start") it's much better, at the trailing edge ("Set end") it's much worse.

Any recommendations for 24/96 => 16/44 => 24/96 for transparency

Reply #49 –
Ah, well I did not use the set start and set end buttons as I recall rather I just switched between samples randomly as it played.

If it is broken though that concerns me. I will try to repeat the result with the new version anyway.
FLAC -> JDS Labs ODAC/O2 -> Sennheiser HD 650 (equalized)